My Current Research

 

Due to the rapid increase in the size and complexity of sensory systems, control of attention is necessary for a robot that

performs a real world and demanding task.

However, attention control and performing the task are tightly interlaced and both should be learned interactively,

particularly when a rational solution is not known at the design time or when it changes over time. Coping with this complex

learning problem from the scratch is impractical in real robotic systems.

 

in my Ph.D. thesis, a framework with three consecutive learning phases is proposed.

In the first phase, a passive demonstration-based learning under the human supervision is performed.

Then, in the second phase, learning evolves to a more active mode and the decision making is transferred to the robot but in a controlled slow speed world.

Finally, in the third phase, the attention control is learned concurrently with motor actions in a normal speed world.

 

The robots mind is composed of a set of tiny agents learning and acting in parallel. One of them fully observes the world, while

the others -partial observers- have access to only a subset of the worlds information. There is also an Attention Control

Learner (ACL) agent.

 

The state of ACL is formed by the belief of those partial observers that are attended to at each state. The problem is

modeled as an optimization process. Besides, due to the fact that the state spaces (perceptual or decision) in this problem are continuous, a Bayesian continuous RL method is employed by observers (who learn in perceptual space) and ACL

(who learns in the decision space). The proposed framework is evaluated on a simulated and a real e-puck robot for miniature highway driving task.

 

The main challenges of the work are:

 

 

Here is my CV